Scholarlink

With the rapid development of smart grid technology, Non-Intrusive Load Monitoring (NILM) plays an increasingly important role in building energy management. To achieve accurate and efficient energy management, this paper proposes an NILM method based on multi-stage Temporal Convolutional Network (TCN) and Informer. Specifically, the temporal convolutional network (TCN) is applied in two aspects: feature extraction and input prediction for the Informer decoder. First, TCN extracts features from aggregated power sequences across multiple time scales. Second, another TCN is employed to predict the target power sequence. Furthermore, the outputs of the two TCNs are fed into the encoder and decoder of the Informer model, respectively. Load disaggregation results are obtained by training the Informer model. The proposed model is compared with three state-of-the-art models and applied to four typical electrical load datasets from UK-DALE. Case studies demonstrate that the proposed TCN-Informer load disaggregation model achieves a reduction of up to 28.26% in mean absolute error (MAE) on the house2 dataset of UK-DALE.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Non-intrusive Disaggregation of Building Loads Using Multi-Stage TCN-Informer

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tengyu Wang,
  • Xing Zhong,
  • Lijie Yang,
  • Tuoyu Chen,
  • Dongdong Zhang

摘要

With the rapid development of smart grid technology, Non-Intrusive Load Monitoring (NILM) plays an increasingly important role in building energy management. To achieve accurate and efficient energy management, this paper proposes an NILM method based on multi-stage Temporal Convolutional Network (TCN) and Informer. Specifically, the temporal convolutional network (TCN) is applied in two aspects: feature extraction and input prediction for the Informer decoder. First, TCN extracts features from aggregated power sequences across multiple time scales. Second, another TCN is employed to predict the target power sequence. Furthermore, the outputs of the two TCNs are fed into the encoder and decoder of the Informer model, respectively. Load disaggregation results are obtained by training the Informer model. The proposed model is compared with three state-of-the-art models and applied to four typical electrical load datasets from UK-DALE. Case studies demonstrate that the proposed TCN-Informer load disaggregation model achieves a reduction of up to 28.26% in mean absolute error (MAE) on the house2 dataset of UK-DALE.